machine-learning-applicationsfor-datacenter-optimization-finalv2.pdf

《machine-learning-applicationsfor-datacenter-optimization-finalv2.pdf》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、MachineLearningApplicationsforDataCenterOptimizationJimGao,GoogleAbstractThemoderndatacenter(DC)isacomplexinteractionofmultiplemechanical,electricalandcontrolssystems.Thesheernumberofpossibleoperatingconfigurationsandnonlinearinterdependenciesmakeitdifficulttounderstandandoptimizeenergyefficiency.

2、WedevelopaneuralnetworkframeworkthatlearnsfromactualoperationsdatatomodelplantperformanceandpredictPUEwithinarangeof0.004+/0.005(meanabsoluteerror+/1standarddeviation),or0.4%errorforaPUEof1.1.ThemodelhasbeenextensivelytestedandvalidatedatGoogleDCs.Theresultsdemonstratethatmachinelearningisaneffect

3、ivewayofleveragingexistingsensordatatomodelDCperformanceandimproveenergyefficiency.1.IntroductionTherapidadoptionofInternetenableddevices,coupledwiththeshiftfromconsumersidecomputingtoSaaSandcloudbasedsystems,isacceleratingthegrowthoflargescaledatacenters(DCs).Drivenbysignificantimprovementsinhard

4、wareaffordabilityandtheexponentialgrowthofBigData,themodernInternetcompanyencompassesawiderangeofcharacteristicsincludingpersonalizeduserexperiencesandminimaldowntime.Meanwhile,popularhostingservicessuchasGoogleCloudPlatformandAmazonWebServiceshavedramaticallyreducedupfrontcapitalandoperatingcosts

5、,allowingcompanieswithsmallerITresourcestoscalequicklyandefficientlyacrossmillionsofusers.ThesetrendshaveresultedintheriseoflargescaleDCsandtheircorrespondingoperationalchallenges.Oneofthemostcomplexchallengesispowermanagement.GrowingenergycostsandenvironmentalresponsibilityhaveplacedtheDCindustry

6、underincreasingpressuretoimproveitsoperationalefficiency.AccordingtoKoomey,DCscomprised1.3%oftheglobalenergyusagein2010[1].Atthisscale,evenrelativelymodestefficiencyimprovementsyieldsignificantcostsavingsandavertmillionsoftonsofcarbonemissions.WhileitiswellknownthatGoogleandothermajorInternetcompa

7、nieshavemadesignificantstridestowardsimprovingtheirDCefficiency,theoverallpaceofPUEreductionhasslowedgivendiminishingreturnsandthelimitationsofexistingcoolingtechnology[2].Furthermore,bestpracticetechniquessuchas